Problem

User bubbles can depend on non-empty cleaned text, so attachment-only turns may lack a normal user-message container. Copy actions can receive raw persisted message text containing attachment or Canvas serialization rather than the human-visible content. Clipboard failure is swallowed, and image attachments are represented mainly by generic file metadata.

User impact

The conversation can lose visual authorship for attachment-only turns, Copy can put internal storage markup on the clipboard, and users receive no explanation when copying fails.

Relevant code

• apps/desktop/src/ui/chat/FeedRow.tsx
• User-message cleaning/attachment parsing
• Message copy action
• Attachment rendering and stored Canvas request format

Implementation outline

• Derive one semantic visible-message model for rendering, copy, accessibility, and export.
• Always render attachment-only turns inside a user-message container.
• Copy the human-visible message and useful attachment names, never raw persistence markup.
• Show safe image thumbnails/previews where supported.
• Surface clipboard failure through inline/toast feedback.
• Handle malformed or legacy stored payloads without leaking serialization.
• Keep the screen-reader author/message structure aligned with #225.

Acceptance criteria

• Text-only, attachment-only, mixed, and Canvas-request turns all have clear user authorship.
• Copy output matches what the user sees.
• Internal attachment/Canvas serialization is never copied unintentionally.
• Copy failure is visible and retryable.
• Supported images have useful previews without unsafe loading.
• Malformed legacy payloads degrade to a readable representation.

Required tests

• Text-only, attachment-only, mixed files, images, and Canvas requests.
• Raw/legacy/malformed stored payloads.
• Clipboard success/failure.
• Accessible message labeling.